0711 深度多视图学习

Deep learning for multi-view data

2. 2011 Multimodal deep learning
   autoencoder-based (splitAE)
   1/3 video only 1/3 audio only, 1/3 both
   encode共享，分别decode
   

• Deep Canonical Correlation Analysis(DCCA)
  two DNNs f and g are used to extract nonlinear features for each view and the canonical correlation between the extracted features f(X) and g(Y) is maximized. U’f is the final mapping

• Deep Canonically Correlated Auto-Encoders
  

• Variational Canonical Correlation Analysis

• On Deep Multi-View Representation Learning: Objectives and Optimization
  

待看:

• Multiview Deep Learning
  Sun S., Mao L., Dong Z., Wu L. (2019) Multiview Deep Learning. In: Multiview Machine Learning. Springer, Singapore

通过构建层次结构获得共享的表示。

• Joint Representation
  • Probabilistic Graphical Models
    Sun S, Liu Q (2018) Multi-view deep gaussian processes. In: Proceedings of the international
    conference on neural information processing
  • Fusion of Networks
    • Split Multiview Autoencoders
    • Representation Fusion

构造具有多个视图的不同表示形式的结构化空间，这给不同视图上的表示形式之间提供约束。

• Complementary Structured Space
  • DCCA
  • DCCAE
    • Correlated Autoencoders (Corr-AE)
    • Minimum-Distance Autoencoders (DistAE)
      

显式地构造不同视图或表示之间的连接或关系，这允许不同视图彼此转换或映射。

• View Mapping
  • Generative Models
    Deep Conditional Generative Adversarial Networks
    Multiview Attention RNN

---

multi-view deep clustering

自表达

• 2018 Deep multimodal subspace clustering networks
  CNN-based autoencoder
  Fig3
  

• Multi-view Deep Subspace Clustering Networks
  auto-encoder
  

• 2019 Deep Adversarial Multi-view Clustering Network(DAMC)
  AE + GAN
  

• Multi-view Generative Adversarial Networks
  BiGAN

• 2017 AAAI Multi-View Clustering via Deep Matrix Factorization
  

• 2018 IJCAI Deep Multi-View Concept Learning

• 2020 AAAI Shared Generative Latent Representation Learning for Multi-View Clustering
  

---

imbalanced deep learning

• data-level /re-sampling --deep CNN
  ROS/ augmentation : SMOTE
  RUS : EasyEnsemble
  hybrid
  Dynamic sampling

1. 从多数类别中删除样本的方法（欠采样，如RUS、NearMiss[11]、ENN[12]、Tomeklink[13]等）
2. 为少数类别生成新样本的方法（过采样，如SMOTE[14]，ADASYN[15]，Borderline-SMOTE[16]等）
3. 结合上述两种方案的混合类方法（过采样+欠采样去噪，如SMOTE+ENN[17]等）

• cost-sensitive

• ensemble learning
  SMOTE+Boosting=SMOTEBoost [21];
  SMOTE+Bagging=SMOTEBagging [22];

–

• Survey on deep learning with class imbalance
• Imbalanced deep learning by minority class incremental rectification
• Learning deep representation for imbalanced classification
• Justin M. Johnson and Taghi M. Khoshgoftaar.Survey on deep learning with class imbalance.Johnson and Khoshgoftaar J Big Data.(2019) 6:27

---

semi-supervised learning

1. Pre-training

• 无标签数据初始化网络,有标签数据微调
  • 无监督预训练
  • 伪有监督预训练
• 有标签数据训练网络, self train
  有标签数据训练一个分类器，然后用这个分类器对无标签数据进行分类，产生伪标签, 把选出来的更可信的无标签样本用来训练分类器

2. 生成式方法
   基于生成式模型的方法。此类方法假设所有数据（无论是否有标记）都是由同一个潜在的模型“生成”的。这个假设使得我们能通过潜在模型的参数将未标记数据与学习目标联系起来，而未标记数据的标记则可看作模型的缺失参数，通常可基于EM算法进行极大似然估计求解。此类方法的区别主要在于生成式模型的假设，不同的模型假设将产生不同的方法。

3. 基于图的半监督算法
   标签传播
   通过构造图结构来寻找训练数据中有标签数据和无标签数据的关系。标签传播过程,从有标签的数据流向无标签数据.

4. 基于分歧的方法
   co-training
   不同角度训练出来的分类器对无标签样本进行分类，再选出认为可信的无标签样本加入训练集中。